Anthropic submitted a confidential S-1 to start the IPO clock, Alphabet raised $80B to extend its compute buildout, and OpenAI landed on AWS. At the model layer, Opus 4.8 tripled GPT-5.5 on a hard reasoning benchmark, Nvidia shipped a physical-AI foundation model, and a US open-weights release tried to catch a Chinese frontier that has already pulled ahead.
The Business Layer of AI Moves
Three stories about how the maturity layer of frontier AI is hardening: who is going public, who is funding compute at scale, and who is no longer locked to one cloud.
- Anthropic files confidential S-1 with SEC for IPO — Anthropic has formally started the IPO process by submitting a draft registration statement, beginning a review window that typically runs three to six months before any road show.
- Alphabet raises $80B in stock to fund AI compute expansion — The equity raise, one of the largest in tech history, includes a $10B placement with Berkshire Hathaway and a $40B at-the-market program beginning in Q3.
- OpenAI’s frontier models and Codex land on AWS — OpenAI ends its Azure-only distribution era, letting AWS enterprises route GPT models through existing IAM, procurement, and billing infrastructure.
The US-China Frontier Tightens
The US closed the most-used workaround for chip exports, and the strongest American open-weights release in a year tried to answer a Chinese frontier that has spent the spring pulling ahead.
- US closes the subsidiary loophole on advanced chip exports to China — The Commerce Department extended export license requirements to any China-headquartered entity regardless of where its subsidiary operates, eliminating the most-used workaround since 2022.
- NVIDIA’s Nemotron 3 Ultra is the strongest US open-weights model — A 550B-parameter MoE model scores 48 on the Artificial Analysis Intelligence Index, outpacing every American open-weights release but still trailing the Chinese frontier.
Frontier Reasoning, Model Welfare, And Multimodal Agents
Three pieces show where the model-capability conversation has moved: a reasoning benchmark that suggests a widening gap, a welfare disclosure that exposes the limits of self-report, and a Chinese release that puts GUI and CLI in one agent loop.
- Opus 4.8 triples GPT-5.5’s score on ARC-AGI-3 — A public benchmark thread puts Anthropic’s flagship in the 60s and OpenAI’s in the 20s on the hardest abstract-reasoning suite yet.
- Anthropic is spending real engineering hours on a question it cannot answer — The model welfare section of the Opus 4.8 system card reveals a serious methodological gap that will matter long before the science catches up.
- Alibaba’s Qwen3.7-Plus aims to unify GUI and CLI in one agent loop — Qwen3.7-Plus merges vision and language into a single agent foundation that claims to drive a browser and a terminal inside the same run.
Physical AI Steps Forward
An open foundation model that outputs robot actions, and an unusually candid interview about what frontier video agents will actually require.
- Nvidia ships Cosmos 3, an open foundation model that outputs robot actions — The first open-weights physical AI model with a native action modality changes how robotics teams approach pretraining, shifting the bottleneck from data to fine-tuning compute.
- What Ethan He actually thinks video agents require — The Cosmos and Grok Image lead says video generation is not the goal; the goal is using generated video as a planning substrate for physical action.
The Builder’s Stack
Three releases that change something concrete in the developer stack, with two competing bets on where RAG goes next and one IDE-native coding model from the JetBrains side of the toolchain.
- Perplexity bets that models should write their own search plans — Search as Code replaces monolithic RAG pipelines with model-generated retrieval logic, outperforming standard approaches on complex multi-step queries.
- Mistral ships Search Toolkit to unify RAG ingestion, retrieval, and eval — The MIT-licensed framework targets the gap LangChain and LlamaIndex leave open: built-in evaluation loops alongside ingestion and retrieval in one interface.
- JetBrains ships Mellum 2, a 12B MoE model built for IDE-native coding — The second-generation coding model targets code completion, tool use, and agentic workflows inside JetBrains IDEs, cutting inference cost through sparse activation.
Today’s Quick Hits
- OpenAI Cookbook shows how to run its models on Amazon Bedrock — A new official guide reduces the integration cost to an afternoon for any team already running workloads on AWS.
- Cursor adds Premium seat and spending controls to Teams plan — Cursor’s June 2026 pricing update separates heavy agent users from standard IDE users for the first time, and gives admins the spending caps they asked for.